Why this exists

Claude.ai enforces a 5-hour rolling usage window. When you hit the limit, the conversation stops. The platform shows a vague progress bar with no token breakdown — you can't see how much each conversation costs, whether a long chat is eating your budget, or exactly when the window resets.

Every other tracker estimates tokens from the DOM. Claude Token Tracker hooks directly into window.fetch() and parses the real SSE stream, capturing token data from the actual API response — including artifacts, dynamic outputs, thinking tokens, and tool use.

Features

How it works

Most token trackers read text from the DOM after rendering. This misses artifacts, dynamic components, thinking tokens, and tool calls — often the heaviest consumers.

Claude Token Tracker uses a three-layer capture strategy:

1. JSON.stringify hook   ← intercepts the full conversation payload before fetch
2. DOM history reader    ← reads all rendered messages as fallback
3. DOM input buffer      ← captures the current input box as last resort

For output, the extension tees the ReadableStream from the SSE response:

// Simplified core mechanic
window.fetch = async function (...args) {
  const response = await originalFetch(...args);
  const [forClaude, forUs] = response.body.tee();
  parseSSEStream(forUs); // we read our copy
  return new Response(forClaude, ...); // Claude gets the original
};

The SSE stream is parsed event by event:

• content_block_delta → accumulates text, thinking, and tool JSON
• message_limit → Claude.ai proprietary event (does not expose token counts)

Since Claude.ai does not expose input_tokens in the stream (unlike the public API), input is estimated using a BPE approximation calibrated against the Claude.ai usage bar.

Token estimation

Natural language  → ~4 chars / token
Code blocks       → ~3.5 chars / token
URLs              → ~3 chars / token
Non-ASCII / emoji → ~2 tokens each
System overhead   → +400 tokens (once per conversation)

Plan limits — empirically calibrated

These limits were derived by measuring the Claude.ai usage bar delta against known input sizes. Anthropic does not publish exact token limits for Claude.ai chat.

Methodology: sent N words of input in a fresh conversation, measured % change on the usage bar, derived tokens-per-percent, extrapolated to 100%. Tested with 1k / 2k / 5k word inputs on Claude Pro.

Installation

Chrome / Edge (Developer Mode)

1. Download or clone this repository
2. Open chrome://extensions or edge://extensions
3. Enable Developer mode (toggle, top-right)
4. Click Load unpacked
5. Select the claude-tracker folder
6. Pin the CT icon to your toolbar

Chrome Web Store release coming soon.

Project structure

claude-tracker/
├── manifest.json          # Manifest V3
├── popup.html             # Extension popup UI
├── popup.js               # Rendering, sparkline, timer, plan switcher
├── src/
│   ├── content.js         # MAIN WORLD — fetch hook, SSE parser, BPE estimator
│   ├── bridge.js          # ISOLATED WORLD — relay between page and background
│   └── background.js      # Service worker — state, alarms, notifications, badge
└── icons/
    ├── icon16.png
    ├── icon48.png
    └── icon128.png

Accuracy

Why not exact? Claude.ai's SSE stream uses a proprietary message_limit event that does not expose input_tokens — unlike the public API at api.anthropic.com. This is a platform constraint, not a technical limitation of the extension. No token tracker for Claude.ai chat can obtain exact counts without server-side access.

Troubleshooting

Tokens not updating after a message? Make sure you're on claude.ai and the extension is enabled. Try refreshing the page once — the fetch hook injects at document_start and needs a fresh page load.

Input showing 0? The stringify capture method may not be firing. The extension will fall back to DOM history reading. Check the small status line at the bottom of the popup.

Numbers seem much higher than expected? Long conversations are expensive — Claude re-reads the full context window on every turn. A 20-message conversation costs significantly more per message than a fresh one. This is working as intended.

Claude.ai updated and tracking broke? Anthropic occasionally updates their frontend. If DOM selectors break, open an issue with your browser version.

Contributing

Pull requests welcome. Key areas:

• Improving DOM selectors for new Claude.ai markup
• Refining the BPE estimator accuracy
• Adding Firefox support (requires world: MAIN polyfill)

License

MIT — see LICENSE

Disclaimer

This extension reads network traffic locally in your browser. No data is sent anywhere. It does not interact with the Claude.ai API — it only observes responses that your browser is already receiving. Use responsibly and in accordance with Anthropic's Terms of Service.